[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20170629154406.44xo7dnw7btn4gpx@redhat.com>
Date: Thu, 29 Jun 2017 11:44:06 -0400
From: Don Zickus <dzickus@...hat.com>
To: Andi Kleen <ak@...ux.intel.com>
Cc: "Liang, Kan" <kan.liang@...el.com>,
Thomas Gleixner <tglx@...utronix.de>,
"linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
"mingo@...nel.org" <mingo@...nel.org>,
"akpm@...ux-foundation.org" <akpm@...ux-foundation.org>,
"babu.moger@...cle.com" <babu.moger@...cle.com>,
"atomlin@...hat.com" <atomlin@...hat.com>,
"prarit@...hat.com" <prarit@...hat.com>,
"torvalds@...ux-foundation.org" <torvalds@...ux-foundation.org>,
"peterz@...radead.org" <peterz@...radead.org>,
"eranian@...gle.com" <eranian@...gle.com>,
"acme@...hat.com" <acme@...hat.com>,
"stable@...r.kernel.org" <stable@...r.kernel.org>
Subject: Re: [PATCH V2] kernel/watchdog: fix spurious hard lockups
On Wed, Jun 28, 2017 at 01:14:04PM -0700, Andi Kleen wrote:
> It can be a useful debugging tool for a specific class of bugs:
> when kernel software is looping forever.
>
> But if that happens does it really matter how many iterations the
> loop does before it is stopped?
>
> Even the current timeout is essentially eternity in CPU time, and 3x
> eternity is still eternity.
That isn't true. We have customers that test the accuracy and file bugs. I
had to write a RHEL whitepaper a number of years ago explaining why the
softlockup took 62 seconds to fire instead of 60.
Customers were changing the watchdog_thresh when the system slowed down to
purposely trigger a panic (which exposed race conditions leading Uli to
redesign the sysctl interface).
I don't feel like explaining to our customers how we regressed on our
watchdog accuracy. It is exhausting, especially if it is a debug feature.
>
> > The hrtimer increase maintains that and just adds a few more
> > interrupts/second.
>
> Interruptions are a big deal for many people.
Yes, and we probably have customers that will complain on that too.
Either solution is a lose/lose. And yes, we will probably get bit by the
false NMI problems on those Intel boxes. This is why I was preferring a
real solution.
The question is, if the real solution is going to take a while, what is the
least sucky solution for now? Or how do we minimize it to a specific class
of Intel boxes.
Cheers,
Don
Powered by blists - more mailing lists